One event per line: STAGE | STATUS | CONTENT. Material claims must include COMPUTED, SOURCED, RECALLED, or ASSUMED.
The audit checks visible procedure. It cannot inspect hidden reasoning.
Fix the earliest dependency first; downstream labels and re-entry are then re-verified before verdict.
The saved conformance auditor parses one stage, status and content record per nonempty line, separated by vertical bars. It records the original line number as the event number, so blank lines can create gaps. The sample has five events: read intent, decompose, execute, a failed self attack and verdict. Execute appears at event three without any allocate stage. The auditor compares the first execute position with the first allocate position and flags skipped allocate at that execute event. The animated path contrasts the observed jump with the documented allocation dependency. This check considers only the first occurrences, rather than enforcing allocation before every later execution. Other graph edges are exported but the analyze function does not generally validate every observed transition against them. The explainer describes visible source behavior, not hidden reasoning or an observed model run. The sample execute content says success probability is forty two percent without an epistemic label. The auditor flags any content containing a digit if it lacks one of four uppercase words: computed, sourced, recalled or assumed. It checks the content field, not the stage or status. The match is case sensitive and does not verify that a source exists or a computation is valid. A version number or date can trigger the same heuristic, while a material number spelled in words can avoid the digit check. Both skipped allocate and unlabeled claim therefore refer to sample event three. The violations are sorted by event number; the earliest display is a line index, not an elapsed timestamp or a proven root cause. The diagram explains a lexical rule and should not be interpreted as model verification. After a self attack with uppercase-normalized status fail, the auditor checks the next event if one exists. It accepts decompose or re derive; the sample proceeds to verdict first and receives invalid re-entry at event five. If the failed self attack is the final event, no next event exists and this branch records no violation. The sample therefore produces three documented violations: two at event three and one at event five. The corrective route is a fixed array of self attack, decompose, allocate, execute, self attack and verdict first, rather than a graph-derived optimal repair. The copied vendor directory is empty, so Graphlib cannot initialize here and Analyze becomes disabled. Sample loading and reset are separate native controls, but a healthy audit or exported analysis remains unproven offline. No model, backend, or hidden-reasoning inspection is performed by this local video.